Swarm intelligent systems (SIS) constitute a significant branch of artificial intelligence, distinguished by their collective intelligence. This emergent property arises from the self-organized behavior of individual agents, ensuring high robustness at the system level. The decision-making process within these swarms is notably complex and nonlinear; it integrates human, machine, and physical elements across multi-dimensional spaces and involves intricate cycles of perception, decision-making, feedback, and optimization. A primary limitation of conventional control algorithms is their dependence on prior knowledge and human expertise, which restricts their capacity for dynamic adaptation and continuous system evolution. As an end-to-end paradigm that unifies perception and action, Reinforcement Learning (RL) offers a promising solution. RL agents optimize their policies through iterative environmental interactions, demonstrating a strong capability for autonomous learning. Inspired by biological systems and AI progress, RL research has evolved from single-agent problems to multi-agent coordination, injecting new vitality into the study of swarm intelligence. Despite this progress, applying RL to swarm systems remains challenging due to issues such as spatiotemporal sensitivity in perception, high individual agent autonomy, complex inter-agent dynamics, and multi-faceted task objectives. This survey provides a systematic review of RL methods applied to intelligent swarm systems, focusing on strategies developed to address the aforementioned challenges. Our analysis is structured along four key dimensions: joint communication, cooperative decision-making, reward feedback, and policy optimization. We also discuss potential future research directions driven by practical requirements.